Back

An interpretable ML model to characterize patient-specific HLA-I antigen presentation

Liang, S.; Jiang, X.; Chiu, Y.; Xu, H.; Kim, K. H.; Lizee, G.; Chen, K.

2023-03-13 bioinformatics
10.1101/2023.03.12.532264 bioRxiv
Show abstract

Personalized immunotherapy holds the promise of revolutionizing cancer prevention and treatment. However, selecting HLA-bound peptide targets that are specific to patient tumors has been challenging due to a lack of patient-specific antigen presentation models. Here, we present epiNB, a white-box, positive-example-only, semi-supervised method based on Naive Bayes formulation, with information content-based feature selection, to achieve accurate modeling using Mass Spectrometry data eluted from mono-allelic cell lines and patient-derived cell lines. In addition to achieving state-of-the-art accuracy, epiNB yields novel insights into the structural properties, such as interactions of peptide positions, that appear important for modeling personalized, tumor-specific antigen presentation. epiNB uses substantially less parameters than neural networks, does not require hyperparameter tweaking and can efficiently train and run on our web portal (https://epinbweb.streamlit.app/) or a regular PC/laptop, making it easily applicable in translational settings.

Matching journals

The top 5 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.